Not a layer, but I believe the plaintiff (the author) would need to prove that it regurgitates their copyrighted work - otherwise it is possibly fair use. OpenAI does not need to prove anything, just defend their position at a reasonable level.

It’s not been decided if training a model on copyrighted works is “okay” or not as far as I know, but I expect it to be so, given that literally everyone does so at this point. It’s not like imagenet is copyright free, many of the images were/are.